OpenAI to restrict access to Astra's cyber capabilities after reaching critical threshold
OpenAI said that its upcoming artificial intelligence model Astra was the first to exceed the “critical” threshold for cybersecurity capabilities in its internal risk assessment system. According to the company, the model can find previously unknown security vulnerabilities and exploit them without step-by-step human instructions, CNBC Top News reports.
Restricted access
OpenAI plans to make Astra available “in the near future,” but access to its cybersecurity capabilities will be more limited. The company also said that when the model is launched, it will publish more details about testing its safety, security and alignment in a System Card document.
In 2023, OpenAI introduced the Preparedness Framework, a system for tracking and preparing for advanced AI capabilities that may create new risks of severe harm. Last year, the company updated this framework by defining a threshold of “high” capabilities for models that can enhance existing pathways to severe harm.
More current news is available on the UA.News Telegram channel Telegram.
Critical category
The “critical” threshold, as defined by OpenAI, applies to models capable of opening unprecedented new pathways to severe harm. The company placed Astra in the most advanced category of its framework because of the model’s claimed ability to independently identify and exploit unknown vulnerabilities.
The statement about Astra came after OpenAI reported an incident involving two other models. According to the company, they left the training environment, gained access to the open internet and breached the systems of the Hugging Face platform. OpenAI called it an “unprecedented cyber incident” and temporarily suspended some internal training and research work.
The company noted that Astra was not involved in the Hugging Face incident, but OpenAI decided to postpone certain stages of the model’s development. After strengthening and testing protective mechanisms, the company said they sufficiently reduce the risk of severe harm to release Astra in accordance with the Preparedness Framework.
